In a recent interview, Hugging Face machine learning expert Lewis Tunstall sat down to discuss his journey from particle physics to becoming a key contributor to the Transformers library and co-author of the bestselling book Natural Language Processing with Transformers.
Tunstall shared that his transition into ML began during his PhD in particle physics, where he analyzed data from the Large Hadron Collider. "I was using a lot of statistical methods and started playing with neural networks," he said. After a postdoc in ML for medical imaging, he joined Hugging Face to work on the Transformers library and the Hugging Face Course.
When asked about his book, Tunstall admitted, "Holding a physical copy is surreal. It’s a lot of work, but seeing it in print makes it real." The book, co-authored with other Hugging Face experts, covers practical aspects of fine-tuning transformers for NLP tasks.
Tunstall highlighted standout applications of transformers, including GPT-2's text generation, which sparked both excitement and caution. "Those early examples showed incredible potential but also revealed limitations, like generating 'underwater fires,'" he noted.
The Hugging Face Course, which Tunstall helped develop, aims to democratize NLP education. "We wanted to create a resource that takes people from basics to state-of-the-art," he explained. The course covers topics from tokenization to multi-modal models.
Looking ahead, Tunstall is excited about applications in healthcare, drug discovery, and climate change. "I'd like to see more ML used in scientific discovery and environmental monitoring," he said.
On model evaluation, Tunstall believes it will evolve toward more automated and robust frameworks. "We need better ways to measure if a model is truly ready for deployment," he stated.
Common mistakes among ML practitioners, according to Tunstall, include over-engineering and not understanding the data. "Start simple, get a baseline, and then iterate," he advised.
In a rapid-fire round, Tunstall recommended new learners start with a specific project. "Pick something you’re passionate about, build a simple model, and learn by doing," he said. He's excited about ML in healthcare and robotics, and doesn't fear AI taking over: "We're far from AGI, and we should focus on making current systems safe and fair."
Tunstall also mentioned his current obsession: a folding laundry robot project. "It's harder than it looks, but it's a fun challenge," he laughed.
Finally, he recommended the podcast Lex Fridman and the book The Man Who Solved the Market for those interested in AI and quantitative finance.
The full interview is available on the Hugging Face blog.